The Past of China's Large Model Capital: 20 Large Model Insiders Walk on the "Life and Death Table"

Tencent News Periscope

2024-05-24 10:23Published on the official account of Beijing Tencent News "Periscope".

Author: Zhang Xiaojun

Producer: Tencent News "Periscope"

The once bustling Texas Hold'em poker game in China's investment circle suddenly fell silent.

"We don't play cards much in the circle anyway." Zhu Xiaohu, managing partner of GSR Ventures, said that his gathering with other investors has changed from playing Depo to tasting red wine to talk about new perspectives on the industry and projects, as well as about stocks in the secondary market.

The real hands are deserted, and the savvy veterans of these games are reaching for the other end. They have played a silent card in the bureau of China's large-scale model industry.

After the advent of ChatGPT, China's large-scale model industry, which will take off in 2023, presents many bizarre landscapes that you have never seen before: a company has been valued at $1 billion in its first round of external financing, and it has just entered the unicorn club without any products; From the first financing, a group of important big people hurriedly "huddled" to place a collective bet; Whether it's Sequoia or Hillhouse, Ali or Tencent, investors who once excluded each other at the Internet table, are happy to sit at a shareholders' meeting this time – you will see that the "playing style" of Chinese venture capital has changed dramatically in the past year or two.

"It's a typical Club Deal." A partner of a first-tier investment institution explained. A "club deal" is a group of investors who form a "club", usually a private equity fund, institutional investor or high-net-worth individual, who jointly invest in the same project. This kind of approach can help spread risk and build up a larger capital plate – in layman's terms, "a bunch of acquaintances with heads and faces, and a bunch of broken silver".

"This is a typical FOMO, which is Fear of Missing Out." Zhu Xiaohu once said. This investor, who used to play Texas with the style of "very tight hands and low entry ratio", did not invest in a large Chinese model company. He commented: "Because everyone is unsure, everyone Club Deal. ”

On the one hand, the myth of mobile Internet wealth making a group of people rich, these high-net-worth individuals have capital, and Chinese VCs have a lot of cash in their hands; On the other hand, due to the confusion about the prospect of AGI, everyone lacks the courage to give it a go, and very few people are willing to spend a lot of money, let alone players who are all-in. "To be honest, Chinese VCs have a lot of money, mainly because everyone is very tight." The managing partner of an investment institution said. So "I'll put a little money in and have a look inside" has become a common choice for most people.

The entrants more or less harbored the mentality that several founders/managing partners of Chinese investment institutions, who did not want to be named, used an almost unanimous expression to describe the psychology of the people in the game - "Everyone wants to see the cards." They said.

This article interviews 20 key entrepreneurs, investors and ecological parties in China's large model bureau, and records the most complete inside story of large model capital in the past year.

Just over a year ago, a superstar entered the game with a heavyweight, and the whole game was lively.

In February 2023, Wang Huiwen (co-founder of Meituan), who retired from Meituan for two years, suddenly announced his return to Jianghu - the establishment of an artificial general intelligence (AGI) company called Light Years away, with the mission of being "China's OpenAI".

In fact, a few days before Wang Huiwen announced his entrepreneurship, his idea jumped violently. "He changed his mind once a day during that time." According to people who were in frequent contact with him at the time, at first, he wanted to invest $50 million, and there were several possibilities in front of him:

First, only as an investor, 50 million US dollars are scattered into 5 projects; Second, he personally participated and acted as an investor, opening up $50 million, buying a "ferry ticket" for himself for $30 million, and buying two "tickets" for others for $20 million, each investing $10 million.

Under the mediation of investors, Wang Huiwen and Li Zhifei (founder and CEO of Mobvoi), who is also in a fanatical mood and also declares to be "China's OpenAI", met for dinner, and the banquet was full of subtle temptation. The people at the dinner table want to see if Wang Huiwen is willing to buy a "luxury warehouse ferry ticket" for herself, and whether she is willing to buy one for this former Google scientist with an AI background. But soon, that didn't happen.

Wang Huiwen made a decision that surprised many people: he not only wanted to go down in person, but also used $50 million to go all in himself.

"Lao Wang bought himself a 'family ticket'." A person close to Wang Huiwen and Li Zhifei at the same time said at the time, "Lao Wang said that it is most reassuring to put it all on himself." ”

Another person close to Wang Huiwen said: "Lao Wang wanted to vote at first, but then he said that it might not work, and after chatting for a while, he felt that he had to do it himself." You don't have to be a CEO to do it yourself. But obviously Lao Wang is not the CEO, and there is a lot of pressure on others to be the CEO. Therefore, the solution beyond light years is that Wang Huiwen is the CEO and is looking for a technical partner.

The founder valued it at $200 million and personally contributed $50 million — and Wang Huiwen, who sat in the "Big Blind" position, raised everyone's betting chips. After that, $50 million was identified as the starting standard for large model companies. If the people behind want to participate in the game, they have to pay the same price.

However, this magnitude of money not only did not deter the comers, but made more people rush to the news.

Light years away, external financing was launched, and at this time, a group of "familiar faces with heads and faces" took to the table.

"It was an exaggeration. The founders are generally one by one investment institutions, and Lao Wang is a roadshow for all investors who are interested and have made offers. The back is basically like a process. A person close to the deal said. This "unified roadshow" gathered by bigwigs was held in the office located light years away in the Sohu Network Building in Beijing.

Cao Yi, founding partner of Source Code Capital, is the most decisive bet. When Wang Huiwen was still undecided, he promised that as long as he entered the game, he was willing to take out a considerable proportion of the fund's funds, just to bet on him. As a result, Source Code became the lead investor in this round of light years away with $60 million.

More than 10 institutions and individuals also participated in this round of financing, which can be described as an all-star lineup: including but not limited to individual investors such as Sequoia China, Wuyuan Capital, Tencent, Zhen Fund, Today Capital, Shunwei Capital, Su Hua (co-founder of Kuaishou), Wang Xing (founder and CEO of Meituan). Since each institution is the first and second in command to negotiate, such as Sequoia China Shen Nanpeng, Today Capital Xu Xin, Zhen Fund Dai Yusen, China Model Capital Bureau is also known as the "number one project", the scene is quite grand.

In fact, this sale goes against the natural instincts of some investors. For example, investors value the solidity of the founding team, and Wang Huiwen and the technical partners who teamed up with him, Yuan Jinhui and Cao Yue, were not familiar with each other in the past; As a technology-based entrepreneur, investors often pay attention to the founder's technical background, while Wang Huiwen does not understand large language model technology; What's more, these veterans are well versed in a rule: "Few projects born with a golden soup spoon in history have been successful." ”

Although the object in front of them was holding a shiny soup spoon, at this moment, they put these thoughts aside. Like Cao Yi's choice, they bet on this person - as a super product manager, serial entrepreneur and the second position of a company with a market value of up to one trillion Hong Kong dollars, Wang Huiwen is too tempting to reassure everyone.

"At that time, financing was much better than light years, and it was the fastest valuation to $1 billion, and it was all invested by investment institutions, tens of millions of dollars." A person close to the financing transactions of a number of large model companies recalled, "It's like Source Code invested tens of millions of dollars, and Suhua invested tens of millions of dollars, both of which are all in attitudes." Other companies shouldn't have that. At the beginning, Ogawa took millions of dollars, and it turned out that there was no big organization. ”

According to a number of people familiar with the matter, due to the extremely high enthusiasm of all parties, the total intended subscription amount of the first round of external financing is as high as 500 million US dollars. But Wang Huiwen didn't ask for that much, and he controlled the financing amount at about 300 million US dollars.

In this way, without a complete technical solution and any products outside Lightyear, the funds were quickly raised, and the post-investment valuation exceeded 1 billion US dollars.

At this moment, there are many founders who are also interested in China's AGI business and are raising funds in the market at a valuation that is light years away. Whether it is Yang Zhilin, the founder and CEO of the Dark Side of the Moon, who started at the same time as Wang Huiwen, or Wang Xiaochuan, the founder and CEO of Baichuan Intelligence, who entered the market one step later than him, their start was not as smooth as light years away. One of the reasons is that "Wang Huiwen has melted too much money".

"You have to accept that it's a reality." Wang Xiaochuan said. He attributed Mr. Wang's starting advantage in part to the latter's business sophistication — "Meituan is a company that crawled out of the dead man's pile." But tactics are not the key to victory. For these "battle-hardened" trading veterans, they are willing to take risks in the world of risk and bet at the first time, and there is only one fundamental reason: Wang Huiwen has helped more people make money, and these people still want to fight big.

"A lot of people want to buy a ticket." An investor on the field said.

Not only at the capital level, Wang Huiwen relied on his personal influence to attract excellent computing power and talents for the first time. Capital, talent and computing power are the three elements of large-scale entrepreneurship (computing power also depends on capital). It was revealed that more than 4,000 cards had been obtained outside the time and year, which was the most sufficient for start-ups in the same period.

If Wang Huiwen is regarded as "one of the people who play cards" on the field, light years away is regarded as a pair of "hands" he gets, and the financing amount is regarded as "chips". Player Wang Huiwen started with the biggest pair of "hands" and deep "chips".

"From February to March, a lot of people were worried about financing, and light years away it was quickly done." However, when the factors of production were in place, "work was about to begin", according to a person close to light-years away, and to everyone's surprise, "he was out of order".

Texas Hold'em is a game of probability, and it is possible to lose with a big hand, and a chance to turn the tables with a small hand.

In many people's memories in the future, Wang Huiwen's "suspected depression" without warning stemmed from the fact that he was too responsible. An employee remembers that since February 2023, in the first three months of its formation light years away, Wang Huiwen has been in a state of excitement and fatigue. The billionaire, who lives not far from the company, arrives at the company every morning at the earliest, leaves late at night, hardly returns home on weekends, and holds meetings until 11 or 2 o'clock in the evening — people have once again seen his efforts to start Meituan, but the difference is that "the body may not be able to bear it".

The time arrives in May. Light-years away, capital, people and cards are in place one after another, but when the founders turn their attention to specific large-scale model businesses, hidden dangers really emerge.

Wang Huiwen proposed a "two-wheel drive" strategy of large-scale model technology and product development. Their division of labor is that Wang Huiwen leads the product team, and co-founders Yuan Jinhui and Cao Yue lead the technical team. Although the technical problems are shared by the joint creation, "as the No. 1 position, Lao Wang must form high-quality decisions". The key decisions about the large model are all difficult and valuable. People outside the industry need to supplement a lot of technical knowledge in a very short period of time to form an effective decision, while for those who are senior in the industry, it may be to listen to instinct and intuition.

For example, "In a few months, what will be the market for computing power?" Is it more reasonable to rent or build your own computing power? This kind of decision-making is extremely costly, and it is possible that "a billion yuan will be invested at once, and half of the financing will be out." "Lao Wang is very responsible", all of which put pressure on him.

Based on the review of many people light years away, Wang Huiwen is very reliable, he wants to think about this matter clearly, but when it comes to the scientific research level, it is far beyond his circle of ability. A big model would cost at least $100 million for a single decision, and he couldn't be sure that the path was the right one. "For Lao Wang, he can't relieve this pressure." One person said, "Unlike the Internet, no one will suspect that Didi can't do it, or that Pinduoduo can't do it, at most they just say that I'm not doing it well." But AI has a lot of scientific research attributes, and that's going to happen — I don't know how to do that. ”

Internally, in May, he began to experience symptoms of not being able to sleep at night and concentrating during the day. "There is nothing special that leads to it, but it can only be said that doing specific business is not the same as the previous state of excitement. It is necessary to do a solid job of specific technology step by step and make products out, and there are endless problems in the middle. "Another point is team management." These large model companies have such a problem, forming a team so quickly, 100 people are mixed together, and these 100 people have never worked together before. Lao Wang is also a person who manages more carefully. The above-mentioned employee said.

In June, news of Wang Huiwen's "suspected depression" came out. There is tension between domestic large models, and "some companies directly rush to the office light years away to poach people". Just when the outside world was stunned, Meituan announced that it would acquire Lightyear Beyond for about 2.065 billion yuan.

"It was a split second." One investor light years away said that from the time they learned that the CEO was sick to the time they returned the money, "the whole process should have been a day." The ending of the story is already familiar to everyone: entrepreneurial partner Wang Xing rescues the field and returns the investors' money to their pockets by acquiring Light Years away.

If Wang Huiwen joined the battle in a high-profile manner, he pushed up the overall valuation and financing amount of China's large-scale model startups by one person; Then his departure indirectly promoted the financing process of several other companies.

The above-mentioned person who knows the financing of many large model companies commented: "Light years are suddenly gone, and other large model companies will soon have money - investment institutions always have to choose - and now, suddenly there is one less strong competitor." ”

The exit in 2023 is light-year-away, and it has become the most surprising "turn" in China's large-scale model game so far. The situation on the field changed rapidly, and after only half a year, the cards were very different.

A person light years away later sighed: "Lao Wang's idea has changed the trajectory of many people. ”

If you compare the entrepreneurial player to the Poker player. Now, there are 6 "players" on top of the hand. Behind each of them stand a group of institutions and individuals with China's ambitions for general artificial intelligence, and they provide the "chips" for this game.

The 6 large-scale model startups on the field are: 1. Zhipu AI, with Zhang Peng as CEO and Tang Jie as chief scientist, 2. Minimax founded by Yan Junjie, 3. Moonshot AI founded by Yang Zhilin, 4. Baichuan Intelligence founded by Wang Xiaochuan, 5. Leaping Star with Jiang Daxin as CEO, 6. Zero One Everything founded by Kai-Fu Lee. (In order of incorporation)

There are two people who started at the same time as light years: one is Yang Zhilin, a post-90s scientist who is an AI researcher, and the other is Wang Xiaochuan, the former CEO of Sogou, who was incubated and spun off by Sohu. The former founded the dark side of the moon in March 2023, and the latter announced the establishment of Baichuan Intelligence with the Sogou team in April. Soon, Wang Xiaochuan was locked as Wang Huiwen's most direct opponent.

Wang Xiaochuan is also an entrepreneur, the same age as Wang Huiwen, both graduated from Tsinghua University, and both are charging on the battlefield of China's Internet - the similarity adds to the flavor of the competition between the two people. Judging from the results, China's star funds almost all bet their chips on light years away at that time.

Just as Lightyears away was raising $1 billion, Wang Xiaochuan had just received $50 million in start-up capital at a valuation of $500 million. "There were no institutional investors in the first round of Baichuan, they were all individuals." An insider said that the source of the funds was Wang Xiaochuan and his circle of friends, such as Chen Danian (co-founder of Shanda Group) and Zhang Bangxin (founder and CEO of Good Future) - the circle of friends provided him with bargaining chips.

"Last year was a state of fast rush." Wang Xiaochuan reviewed and said that due to the late entry into the game, "we must be a late one in the competition for talents and funds." Therefore, when entering the market, he adopted the "follow strategy" in the business war. "Not advanced." "There's at least a simple logic in business – if the opponent is more mature than you, you just follow along; But in your areas of strength, you play better. In his view, his main battlefield is technology, and "in essence, Meituan's success is not driven by technology."

According to Wang Xiaochuan, "In the first round, there were first-line investment funds that were willing to invest in us, and even thought that our price was very good, but because we benchmarked and learned to take the money from the circle of friends, we didn't take VC money." My choice is relatively simple, that is, to face the entrance and follow first. ”

In terms of the pace of fundraising, Baichuan is one step slower than light years away. It only completed its second financing of $300 million in October 2023, and it is still rare to see VC - but this time, it presents the characteristics of "war investment is big" - investors include more than a dozen, Tencent, Alibaba, Lei Jun's Xiaomi/Jinshan/Shunwei Capital, Good Future These strategic investors come together, in addition to Tsinghua Holdings, Shenzhen Capital and Red Dot China. Wang Xiaochuan said that unlike Wang Huiwen, who helped VC make a lot of money through Meituan's listing, it was large enterprises that he really made money. "The companies I have made money for are Sohu and Alibaba, including selling Sogou to Tencent."

So, the criteria by which capital measures these two entrepreneurs is very simple: Who has helped more people earn more money?

Just when Wang Xiaochuan and Wang Huiwen were going head-to-head, the dark side of the moon founded by Yang Zhilin was just an "underwater project" filled with "non-consensus", and the valuation was the cheapest.

Zhilin Yang graduated from Tsinghua University and Carnegie Mellon University, and his academic and work are related to large language model technology. His style is more Silicon Valley, and the company only makes to C products (Kimi smart assistant). Although Yang is somewhat famous in the AI circle, he is unknown compared to a group of well-known entrepreneurs, scientists, and executives.

In fact, at the beginning, many people preferred to see him as a CTO-style talent - at the beginning of 2023, someone pulled Yang Zhilin to talk to Wang Huiwen to see if Yang was willing to be the latter's technical partner, but he did not agree.

"At that point in time, I felt that he was suitable for the consensus of right and wrong." A dark-side investor said that the more common view is that Yang is young and his last startup is not very famous, "and many people think that in China, you have good technology and are useful"?

"We dug him out of the water." Another investor on the dark side of the moon said, "At that time, when I was talking in this circle, I would hear him (Yang Zhilin) from time to time - some people said that he was very good, and some people were a little unconvinced. When I heard it for the third time, I thought this person was important. ”

In Yang's projects, Sequoia China has the sharpest sense of smell, which is a shareholder of Yang's last start-up. At the beginning of 2023, Sequoia realized that Yang might be a suitable candidate for entrepreneurship in the large-scale model arena, and the two sides hit it off. But on the question of whether Sequoia is a single vote or a number of votes, there have been several twists and turns.

A person close to the deal said: "At first, Sequoia reported a valuation of about $100 million, but Zhilin found that $100 million was not enough to raise $30 million, and it was impossible to dilute it by 30%, and later raised it to a valuation of $200 million, for example." "The deal is ready to close.

What I didn't expect was that today's capital suddenly "killed" in.

"They 'grabbed land', they were very ferocious, and their valuations were getting higher and higher." According to multiple people familiar with the matter, they pushed the valuation to $300 million in one fell swoop. In order to grab the deal, Today's Capital tried to persuade Yang Zhilin: You should take more, you should take both. A few more weeks passed, and "FunPlus and FunPlus were the first round of last-minute progression."

"Today's valuations have been raised for everyone, except Sequoia." Several insiders revealed. In the end, the deal was closed in two different prices, with Sequoia valued at $200 million and others at $300 million, for a total of $60 million in initial financing. Therefore, when the door of the unicorn was kicked open as soon as it was light years away, and Baichuan Intelligent implemented the follow-up plan at a valuation of $500 million, the dark side of the moon got the bargaining chips at a valuation of $2 to $300 million.

For them, Yang recalled, the key to this round of financing was the time window. In February 2023, before the company was established, he launched his first financing while researching in the United States. "If it is delayed until April, there is basically no chance, and the real window is one month." At this time, he made a calculation, "I think I will get 100 million US dollars in at least a few months."

Excluding the three players who played at the same time, Minimax and Zhipu AI are large model companies that were established earlier.

Zhipu AI can be described as "rooted and sprouted", born in 2019 from the Knowledge Engineering Laboratory of the Department of Computer Science of Tsinghua University, and started an independent business. Early investors include Qiming Venture Capital and Legend Capital. Overall, Zhipu AI is RMB-style. Zhang Peng said that the large model has given this generation of AI companies an opportunity: to be a "technology platform company" rather than a "vertical company that stands tall in the sky".

Standing at the beginning of 2023, the only one that can compete with light-years away in financing is Minimax. Founded in 2021 with a core team from SenseTime, this start-up already has relatively mature technologies and products at this time, and has come to the third round of financing.

Born in 1989, Yan Junjie graduated from the Chinese Academy of Sciences, started as an intern at SenseTime, and worked his way up to vice president. He was previously the CTO of SenseTime's Smart City Business Group, which was one of the company's key revenue sources. "He was the first entrepreneur to talk to me about the Foundation Model." Chen Yu, a partner at Yunqi Capital and an angel investor at Minimax, said that the two sides contacted as early as January 2021. "They now have a 'foundation model' that they understand that for a company to grow, it needs to find the right 'nail'."

In his opinion, Yan Junjie is "a very rounded person". Chen Yu said that in their earliest contact with Yan, they discussed not only technology but also commercialization — discussing projects such as digital humans, smart cities and 3D reconstructions in museums.

According to a number of its shareholders, Minimax has completed two rounds of financing earlier: the first round is around the National Day in 2021, and it has completed a $20 million financing at a valuation of $200 million, with investors such as Hillhouse, miHoYo, IDG and Yunqi Capital. At that time, they met Zhang Lei (founder and CEO of Hillhouse) and Li Liang (founding partner of Hillhouse Investment), and they took pictures directly. The price set by Hillhouse can only be followed by others. ”

The second round was six months later, and in May 2022, it completed a $50 million financing at a valuation of $500 million, and the investor was the above-mentioned old shareholder Jiamingshi Capital.

So, after ChatGPT ignited a global craze, Minimax soon became the second Chinese player to join the unicorn array. From the end of 2022 to April 2023, it quickly raised more than $200 million at a pre-investment valuation of about $900 million, with a post-investment valuation of nearly $1.2 billion. In addition to the old shareholders, in the third round of financing, investors lined up to enter the market include Oasis Capital, Tencent, Xiaohongshu and Shunwei Capital.

Compared with the dark side of the moon, Zhipu AI is biased towards B, and Minimax is more like a "hybrid pie" - both to C products and to B; Both domestic and overseas. In fact, Minimax's products have undergone several adjustments: from the very beginning of the digital human to the AI social product Glow, and then it served as an app factory, launching products such as Conch AI, Talkie, and Hoshino. One long-time shareholder expressed surprise at the development, saying, "Surprisingly good." ”

In addition, the other two contestants are Stepleap Star and Zero One Things, which will be formed in April and May 2023, respectively.

As of the first half of 2023, all 6 "players" are in place. They started at different times, and the founders and the company's genes are very different, but they all caught up with the golden time when capital was willing to invest intensively in China's large-scale model business. Of course, a large part of this is due to the unprecedented fear of capital - the FOMO sentiment of fear of missing out on technological change; And the thirst for greed – AI may be a more majestic business giant than the mobile Internet. For those who came later than them, such a window of time and a psychological window both disappeared.

However, at the moment, the winning rate and odds of the 6 "players" at the Chinese large-scale model table are still quite vague.

"Everybody can't see clearly." A founding partner of a first-tier investment institution said that as a result, many investors made a very consistent choice: "Everyone puts $10 million to tens of millions of dollars and looks at a card." ”

The large model has brought a very special capital feast to China.

On the one hand, it was born out of a series of cramps and cracks—more tense international relations, more limited funds, narrower exit routes, more prudent pitchers, and fewer entrepreneurs; On the other hand, it underpins a sufficiently ambitious vision of science and technology. As a result, the people on the field were excited and vulnerable.

The excitement is that this may be the only capital story that meets the aesthetics of the dollar VC at the moment. "Thank goodness, otherwise I don't think I have anything to do." Here's how one investor described the mood after the AI wave exploded. "Since autonomous driving and O2O, there are very few such rounds of financing of the order of several hundred million dollars." Another investor said.

The fragility lies in the fact that the risks faced by large model companies are too great, once the scientific research fails, or the giants crush it, so much money will not even be able to see the splash.

In this context, the VC group mentality has changed. A fund partner described their psychology as follows: "Everyone wants to invest, but they are worried that they will lose too much money", so the approach becomes: keep close contact with the entrepreneurs, invest a little money, and look at the cards first. "A lot of funds think, I can only invest $100 million, so wait and see. I'd rather vote more expensively, I want a certainty. This has led to the fact that for a long time, investment institutions have bet on chips with an incredible attitude of "making up members", "you are 10 million, I am 10 million, and he is 10 million".

Even though there are similarities in mentality, the timing, strategy and price of each institution are very different.

According to a number of core investors, Sequoia is the most ferocious of VCs. In addition to Lightyear, it has bet on 4 of the 6 on the field today, including Dark Side of the Moon, Spectrum AI, Minimax, and Step Star.

Sequoia's shot is concentrated in the first half of 2023. "They're a way to play that can't be missed." The industry classifies this investment style as "track coverage". At that time, Shen Nanpeng, the founding and managing partner of Sequoia, would personally meet with the large-scale model entrepreneurs. While other funds were still swinging around a few startups, Sequoia put money in multiple startup bureaus.

As mentioned above, Sequoia is a natural angel investor in the dark side of the moon; But in Minimax's game, it didn't get into that easily — on the board of directors of this company sat Sequoia's old rival, Hillhouse.

"It's not that Hillhouse won't let it in, it's just that it doesn't want it to enter so early." A person close to Minimax said. Dramatically, Sequoia's entry into the game was caused by an accident.

In mid-2023, ByteDance will go on the field to look for large-scale model targets, and it is thinking about the possibility of investment at this time. "It had two in mind, one was Minimax and one was the Leap Star." Subsequently, Minimax opened a separate round for Byte at a pre-investment valuation of about $1.6 billion, which was supposed to be Minimax's fourth round of financing. "At least we've seen the SPA [Share Subscription Agreement]." Said a Minimax investor. But what they didn't expect was that Byte repented on the spot - they decided to "either do it themselves or buy a company".

"In the fourth round, it was the bytes that ruined the SPA, and Sequoia was on top." He said.

So, Sequoia squeezed into Minimax's game at a relatively expensive price. "A lot of it is not clear at that stage. No one can say at that point in time, who will do it, and can only make the best choice at the moment. When every possibility is possible, you can only bet on the ones with the highest probabilities. A person on the field analyzed, "What if the Minimax is good?" ”

Internally, the Sequoia mulching strategy is related to its management mechanism. They have a large number of investors, and they need a partner sponsor (support) to invest in each project. In Sequoia, the four projects are scattered under the names of different partners and teams - according to several people close to Sequoia, the dark side of the moon is Zheng Qingsheng, the Zhipu AI is Guo Shanshan, the MiniMax is Zheng Qingsheng, and the step star is Sun Qian - "Neo (Shen Nanpeng) is also listening to the Ming. An industry insider said.

In fact, although many institutions want to "watch the cards", the "card viewing fees" that each can afford are completely different.

When it comes to "card fees", a fund partner calculates that if Sequoia (USD) has about 9 billion ($9 billion) in hand, $100 million is about 1% of its fund size, "so it is acceptable to give 0.5 points and $50 million to look at the cards." And some funds have a management scale of hundreds of millions of dollars, in this case, "losing 1%, is millions of dollars, which is the limit of the affordable viewing fee". In that sense, the millions of dollars these institutions are worth about $100 million for Sequoias—the depth of their pockets determines the amount of fees.

However, Hillhouse, which is comparable to Sequoia, did not adopt the strategy of "casting a wide net", but only made two shots. Among them, Hillhouse re-injected Minimax from the angel wheel, and also put money in Zhipu AI.

The contrast between Sequoia and Hillhouse's investment strategies may be partly due to their different conjectures about the end of the market: an industry veteran has asked whether China's big model will be a winner-takes-all industry.

The first possibility is that it will. If so, you can't miss it anyway, because "if you miss it, your fund's income is very different from others". The second possibility is that no, the market share is not concentrated. "At this time, whether you invest in both, or one or two of them, there is no significant difference in returns. The person who wins the first place in the bet may earn 20% more than you, and that's it. ”

"Now the question is, you can't tell if it's the first or the second." The person went on to say that Hillhouse may be inclined to the second type - domestic large-scale model companies are still in the stage of copying OpenAI, and "there is no magic trick for them to climb to the sky in one step".

After reading the "two deepest pockets" of Chinese VCs, let's look at another wave of funds. They have different ranges and pockets, but they all have a common choice: to shoot earlier or to shoot early.

"In large models, it is better to vote early than to vote well." An early investor in a large model company said. (Even Sequoia and Hillhouse have adopted this strategy in stages.) ）

Capital Today is an active but low-key fund that uses this strategy, and it has bet on Light Years Away, Dark Side of the Moon, and Spectrum AI. Among them, the first round was bet on Light Years Away and on the Dark Side of the Moon. Zhen Fund is also betting on Lightyears Away and the Dark Side of the Moon in the first round. It is worth mentioning that although today's capital has "a lot of chips", it will stop in the second half of 2023 out of risk trade-offs, and has not added any large model companies - "Their guiding ideology is not to invest." Industry insiders.

"For large-scale model investment, the best activity window is one month (February to March 2023), and the valuation will quickly become much more expensive after that month." Another early investor said, "This kind of long-term track (artificial intelligence, autonomous driving) is risky in the first, third or even fifth year, so you might as well invest in it early." In his view, "the risk has not changed fundamentally because of the increase in valuations - you put money early to buy the same chips with less money; Spend more money on chips in the back without any reduction in risk. ”

Traders in the primary market are high-risk bearers, and they always calculate the winning rate and odds in their minds, and choose the timing and trading strategy accordingly.

There are also some unusual shots on the field. Founded by former Sequoia China partner Cao Xi, Monolith is a cutting-edge institution that has only been established for two years. Cao Xi tends to choose only one of the most promising companies to bet heavily.

Cao Xi later recalled that at the beginning of 2023, he got investment opportunities from several start-up large model companies. His idea was, "Make sure you're eligible to play at the table first, and then talk about which one to choose if you have limited funds." Cao Xi pondered for a whole month and made a decision that was extremely difficult at the time: when the market was enthusiastic, he did not put money in several star projects.

Then, he found the younger Yang Zhilin and invested in the first, second, and third rounds of the dark side of the moon one after another. So far, the new fund has risen to the second largest VC institutional shareholder (the first is Sequoia) with an investment of about $38 million.

Investments are similar to Poker. In both activities, the "player" needs to calculate the probability from the cards in hand, the limited information on the field, and the behavior and psychology of the participants, and then make betting decisions. As the information on the court gradually becomes fuller, the chips become heavier and heavier, "Isn't this like every round of investment"? In addition, the "player" also needs to manage the chips and distribute them reasonably, so as not to lose a hand and incur excessive losses.

In Texas Hold'em, players will summarize the Poker style as: Tight-Aggressive, Tight-Passive, Loose-Aggressive, and Loose-Passive. Cao Xi said that he didn't play much in Texas, mainly to make friends, but he liked to take Texas as an example. If it is an investment, the style he hopes to maintain is "definitely tight" - "don't invest so much, there is a person who thinks it may have a high probability of winning, and place a heavier bet." ”

In the game, aggressive players have a strict hand selection, only play high-quality starting hands, and once they decide to enter the game, they will actively bet and raise in search of the initiative.

However, Cao Xi emphasized that investment is a serious business practice, and the use of Texas Hold'em as a metaphor is only for visual expression and easy to understand. In addition, there is a fundamental difference between the two. "Texas, the cards you threw away, they were thrown away. Invest in every card you fold in the past, and you can buy it back later - you didn't vote in round A, you can vote in round B; I didn't vote in the B round, but I can vote in the C round. He said, "So, Swallow your ego, admit if you're wrong, admit it, and then go back."

In addition, there are some investment institutions on the field that have been reluctant to make a move before, and they want to wait for a clearer hand.

For example, Matrix Partners has no bets in 2023. Several people close to Jingwei said that they considered that putting tens of millions of dollars to look at cards was not their favorite way to play, and they hoped to wait for the situation to be clearer, and firmly bet on one family - "invest 50 million or more in the end of one". Jingwei concluded from the autonomous driving track that we should first think clearly about "who is the ideal car" before making a move, so that we can cover the upstream and downstream of the industrial chain with an ontology enterprise as the starting point.

However, in the first half of 2024, we exclusively learned from many parties that Jingwei finally made a move - it put its money on Minimax, "quite a heavy bet".

As soon as he entered the game, he re-invested his chips in Light-Year Away and led the company's Source Code Capital, which passively withdrew from the game because of the exit Light-Year Away. After that, it did not immediately re-bet. Also in the first half of 2024, Source Code put money on the dark side of the moon.

Also newly preparing to sell is Gao Rong Venture Capital, which will most likely put its money on the dark side of the moon after holding the currency for more than 1 year (the deal has not yet been officially signed).

An interesting observation is that in this round of large model waves, in addition to the No. 1 position of the fund, there are many post-80s investors. They have either been promoted to managing partner or become partners. For example, Jingwei Wang Huadong, Yunqi Chen Yu, Qiming Zhou Zhifeng, Lisi Cao Xi, etc. They are more ambitious and eager to take an investment career to the highest level with an adventure.

However, on the sidelines, there is a rather special investor. Here's what he said of himself: "I don't play much now, so I'm not very willing to look at cards now. Like I have a very tight hand, and I am very cautious when looking at the cards hahaha. ”

He commented on his peers: "It doesn't matter to them anyway, they have a lot of money, to be honest." Everyone puts some money to look at the cards, and they all have this mentality. ”

Therefore, Zhu Xiaohu did not want to pay attention to the hustle and bustle of the game, he neither looked at the cards nor bet, shook his head and walked away.

The large-scale model industry has rewritten the rules of the game for China's venture capital transactions.

Different from the scene of Silicon Valley VC's "almost full-line miss large-scale model investment" and "large-scale model is all supported by giants", Chinese VCs have actively taken action with giants in the past year to support domestic large-scale model companies on the road from 0 to 1. The reason why they are willing to do it is simple - "because the United States has a benchmark." Li Guangmi, founder and CEO of Shixiang, said, "We are good at problem solving. ”

However, in this game, the "player's" style of play has changed in a consistent way.

On the one hand, this is constrained by the global economic downturn, and more importantly, the amount of capital consumed by AI is too large - hundreds of millions to billions of dollars in a single financing, and a large part of it is used for high-risk scientific research. As a result, VC's past investment strategy of "growing big with a small one" and "beating the drum from the Early Stage to the Growth Stage" has failed.

"The most classic VC logic is that each check is less than $5 million, and I only write ten for $50 million, although each has a high risk, but there are always two or three out of ten, which is a huge leverage." Wang Mengqiu, founding partner of Crystal Stream Capital, said, "Now the deal of AI, because I have a large single deal check size, it is a Club Deal from day one." This means that these large model companies will not only have one or two shareholders standing on the investment list, but will follow a large number - "everyone is rushing to the top" and "everyone is diversifying risks".

The distribution of bankrolls changes the rules of hand concealment. "Because everyone's money is only enough to spread out to a few, it will become, a project itself has fewer competitors in this track, and in the end there may be 2-3 or even only 1 left - everyone will not feel that 1 will not be able to succeed - then, I have a 1/2, 1/3 chance (to win)." Wang Mengqiu said.

As a result, you will see that on the large-scale model track, 6 startups have emerged to compete. For each of their financing structures, the Club Deal phenomenon is remarkable. Take the dark side of the month with the highest financing amount as an example, it has completed three rounds of financing so far:

The first round is in the first half of 2023, it has raised $60 million at a valuation of $200 million to $300 million, and five investors have jointly sold: Sequoia China, Capital Today, Zhen Fund, Lisi Capital and FunPlus.

The second round is in the second half of 2023, when it closed about $200 million in financing and was valued at about $700 million after the investment. This round was led by Meituan Dragon Ball, Alibaba and Ant, and followed by many investors, including Suhua, Lisi Capital, Jiu'an Medical, Lanchi Venture Capital, Xianghe Capital, etc.

At the end of the year, the Dark Side of the Moon opened an additional small round (equivalent to the + round of the second round). "There are two companies coming in, Wuyuan and Yunjiu, which is not much money." A person familiar with the matter said that they were investors with a "look at the cards" mentality at that time.

Until the end of 2023, the dark side of the moon has raised less than $300 million in total, lagging behind in the capital competition. During the Chinese New Year, a super deal at a special point in time helped them turn the tide of the battle.

On the night of the Chinese New Year's Eve in 2024, tense business negotiations continued into the last few seconds, the deadline for the signing of the third round of funding documents on the dark side of the moon. "Because there are many old shareholders, there is no opinion at about 00:00 on the first day of the Lunar New Year, and that day is considered a close day (the end of the trading day)." At this time, the dark side of the moon completed more than 800 million US dollars in financing, led by Ali and followed by Lisi, with a post-investment valuation of more than 2.5 billion US dollars. However, the third round is not over here.

"It's only part of the deal that closes before the Chinese New Year." Several insiders revealed, "It also has a clause that you can get another $200 million in the next 60 days." It is understood that in the end, the company's third round of financing totaled slightly more than $1 billion, which is the highest single fundraising of China's large model so far; It is currently in the process of raising its fourth round at a pre-investment valuation of $3 billion.

In just one year, the company has risen from obscurity, its valuation has increased 15 times, and it has gone from being one of the players with the least "chips" to having the most.

On the field, Alibaba is the biggest variable since the start of 2024. We have learned from a combination of many parties that the negotiations between the dark side of the moon and Ali took a short time. In order to facilitate this deal, the two parties made some special designs on the transaction:

1. A part of Alibaba's investment is placed in a condominium account for the purchase of Alibaba Cloud services; 2. The founding team of the dark side of the moon has super votes; 3. Ali occupies a seat on the board of directors, but does not have veto power (Veto) on financial terms such as financing and mergers and acquisitions - which means that Ali cannot prevent other investors from coming in.

Since then, Ali's ambitions have not stopped at the dark side of the moon. A few days later, it further funded Minimax. This is Minimax's fifth round of financing, and it has completed more than $600 million in financing, of which Alibaba led about $400 million, with a post-investment valuation of more than $2.5 billion. According to its shareholders, there are no restrictive terms in this financing.

A strategic investor commented on the situation in the first half of 2014: "If it does not fall light years away, it will be unique to itself in one echelon up, and others in one echelon." Today's echelon is gone, and now there are one or two more outstandings in other people's echelons - and this outstanding, which Ali evaluated as a referee. ”

"You've seen it now." Chen Yu said, "If you can't raise a lot of money, you won't have computing power, and you won't be able to make the best model - this is the status quo of large models." "However, the further back you go, the more powerless VC becomes, and China's large-scale model business depends on the hands of giants.

Among the giants, Alibaba is the most aggressive but late, investing in 5 of the 6 players on the field - Dark Side of the Moon, Minimax, Zhipu AI, Baichuan Intelligence and Zero One Everything.

According to the analysis of a number of fund GPs (managing partners), Alibaba's radical further development is related to the appointment of Wu Yongming, who has an investment background, and Alibaba's emphasis on cloud business. "Wu Ma (CEO of Alibaba Group) is a new official, more decisive and more willing to act." In addition, the reality of Ali's urgent need for change is also the source of motivation to force it - a GP said that Ali's ticket to the mobile Internet has not been completely kept, and Pinduoduo and Douyin are fiercely attacking the hinterland of e-commerce, "AI is a ticket to continue life".

In addition to Lightyear, Tencent has invested in Minimax, Zhipu AI and Baichuan Intelligence, among others. "Tencent is going to take a look first." The CEO of a large model company that has received investment from Alibaba, Tencent, and Lei Jun at the same time said, "They made it very clear that we would make the model ourselves, but we were also worried that we would not do well and lose this opportunity, so we also invested." ”

Lei Jun and Meituan are closer to Tencent in terms of mentality. Lei Jun's system (including Xiaomi, Kingsoft, Shunwei Capital, etc.) bet on Light Years Away, Baichuan Intelligence, Minimax, Zhipu AI, and the Dark Side of the Moon. In addition to the acquisition of Lightyear, Meituan has divided its investment into two teams - Meituan Zhan has bet on Zhipu AI, and Meituan Dragon Ball has bet on the dark side of the moon. However, according to people close to Meituan's top management, Wang Xing's mentality is biased towards learning and observation.

Unlike the two-pronged strategy of "self-research + investment" of these companies, Byte and Baidu are betting on their own at this stage.

There are also many well-known entrepreneurs who are also actively reaching out to large model companies. Either out of strategic offense, or out of strategic defense, or simply want to be part of the AI revolution. They "hover" over the table, thinking about strategy and waiting for the best moment.

You will see that China's top entrepreneurs, individual investors and institutions are gathered in the top game of this grand model. They put aside past grievances and collectively contributed to a grouping rather than an exclusive strategy. Club Deal also embodies the pragmatism of today's Chinese venture capital.

"It's hard for people to make money now, and they can't afford to lose much, so they only invest in the best. But the best people are very open-minded people, and the valuation becomes very high. A managing partner of an investment institution observed, "This becomes, everybody flies to quality." (Fly to quality refers to increased market uncertainty or increased risk aversion, and investors tend to buy higher-quality, lower-risk assets.) ）

Wang Mengqiu even felt that "the concept of VC (Venture Capital) is no longer valid, we are not Venture, we want to try not to take risks." She said that in the past, VCs relied on non-consensus to make money, but now it is not. Because there is limited money in the market, everyone must believe in one direction to be willing to continue to invest money. As a result, VCs rely on consensus to make money.

This is a special story of China's capital at a special time - it is silent and noisy, languishing and generous, united and shrewd, and full of contradictions.

It is worth noting that today these "players" still have a "small card" in their hands: their model capabilities have not yet been proven to exceed the level of overseas open source models.

At the same time, the $1 billion financing amount is still a drop in the bucket in terms of global model competition. According to statistics, OpenAI has raised a total of more than $10 billion, and its latest valuation is over $80 billion, an order of magnitude more than that of Chinese companies.

In the past year, while China has grown 6 large-scale model unicorns, there have been deep alliances of multiple giants + model companies in the United States. For example, Microsoft + OpenAI, Amazon + Anthropic, Google in a class of its own, Tesla + xAI. The industry believes that in order to obtain relatively unlimited capital and resources, the binding of model companies to giants is one of the foreseeable ways out.

In this regard, the mentality of some Chinese investors has changed slightly. On the one hand, the giants' largesse has increased the sense of security of VCs, and on the other hand, they feel lost in the fate of model companies that may be swallowed up by giants. Several investors made similar statements: they wanted to bet on "a company with great risks and great returns", and they wanted to capture a giant unicorn worth hundreds of billions of dollars or more.

But what if, a startup ends up selling to a giant for $5 billion or even $10 billion?

"I can understand that, too." One investor said, "For me, it was a very successful failure, not a failed success. ”

In the large model of playing cards, the complex is the model, but also the game of human nature.

- End-

Information update: Zhu Xiaohu said that he has recently started playing cards again, and he is used to playing with investors in the primary and secondary markets. He said: "Now I play more cards, eggs, haha." ”

Special note: Li Zhifei had the idea of making a large model in February 2023, but in April of the same year, his mind changed, and he decided to continue to make a door-to-door question, and the company was listed on the Hong Kong Stock Exchange as the "first AIGC stock" not long ago.

(The source needs to be indicated for content citation.) ）

"Review of China's Large Model Series" is recommended reading

The dark side of the moon Yang Zhilin reviewed the year of large-scale model entrepreneurship: moving forward to the endless and unknown snow-capped mountains

Zhu Xiaohu tells a story about Chinese realism AIGC

Wang Xiaochuan: The third possibility of China's AGI

AGI ushered in a big infrastructure construction tide: with 5 hours of electricity in Shanghai, GPT-5 can only be trained for more than three months

View original image 527K

• 

• 

• 

• 

• 

• 

•